AFRINT 3: Ghana Micro Report
GHANA:
Household–Level Farm–Nonfarm Linkages and Household Welfare
Implications
Fred M. Dzanku Institute of Statistical Social & Economic Research
University of Ghana [email protected]
&
Daniel B. Sarpong Department of agricultural economics & Agribusiness
University of Ghana [email protected]
August 2014
AFRINT 3: Ghana Micro Report
1
Executive Summary
This report has examined linkages between the farm and nonfarm sectors using data collected
from a sample of households in eight villages over the period 2002-2013. The welfare
implications of nonfarm participation and income have also been analysed using two welfare
indicators at the household level: a composite wealth index and food security. The regression
analyses were based on the last two rounds of surveys (2008 and 2013) since these contained
the relevant income data which is central to the theme of this report. More detailed analysis is
provided using the most recent round survey data because even more detailed income data
was collected. Indeed, the analysis using the most recent data is where this report contributes
to the existing literature on farm-nonfarm-linkages.
Over the last two waves of the surveys, overall farm size increased significantly, from an
average of 2 ha in 2008 to 2.6 ha in 2013. For the three staple crops studies in detail (maize,
sorghum and rice) farm sizes increased significantly for maize and rice but not sorghum.1 Maize
yield in the Eastern region remained largely unchanged but all three staple crops experienced
significant yield increases in the Upper East region between 2008 and 2013.
This report is in part concerned with examining the effect of nonfarm income on farm output
through its effect on farm input use. On farm input use, it is observed that the proportion of
farmers using improved seeds declined significantly between the last two waves of the panel.
Significantly more farmers were using inorganic fertilisers (47% in 2008 compared with 37% in
2013) but the quantities being used remained largely unchanged. The proportion of farmers
using hired farm labour remained unchanged between the two periods. Significantly fewer
farmers had contact with agricultural extension agents (57% of farmers in 2008 compared with
50% in 2013).
Significantly fewer maize producing households sold their output in 2013 compared to 2008
(about a 19 percentage point difference), and conditional on selling, the average share put on
the market also declined (from 66% in 2008 to 57% in 2013). The opposite is observed for rice:
output market participation significantly increased by about 19 percentage points between
1 Unless otherwise specified, in this report, the word significant or significance is used in the statistical sense.
AFRINT 3: Ghana Micro Report
2
2008 and 2013; conditional sale shares increased by 11 percentage points (from 41% of output
in 2008 to 52% in 2013).
Turning to incomes and the nonfarm sector in particular, it is first observed, as expected, that
crop income accounted for the largest share of household total cash income in the higher agro-
potential Eastern region (77% in 2008 and 64% in 2013, representing a significant decline over
the period). Crop incomes were less important in the Upper East, accounting for 19% and 29%
of household income in 2008 and 2013, respectively (the increase over the period is
significant). Over the entire sample, crop incomes represent 46% of household total cash
income. Participation in nonfarm income increased significantly by 24 percentage points over
time (57% in 2008 and 82% in 2013). Similarly, the share of household income derived from
nonfarm sources increased by approximately 13 percentage points (30% in 2008 compared
with 43% in 2013). Significant regional differences exist: over the two periods, average nonfarm
income share among Upper East region households was twice the share among Eastern region
households (23% for Eastern region households versus 47% for the Upper East).
The gender disaggregated data from the 2013 survey provide some important results. First, it is
observed that using the detailed income data increases the estimated overall nonfarm
participation rate by approximately seven percentage points, the difference being highly
significant. This means that nonfarm participation was significantly underestimated when not
collecting detailed gender disaggregated information. The intra-household descriptive analysis
shows that women have significantly fewer number of income sources, earn lower incomes,
have a higher nonfarm income share (mostly from nonfarm self-employment income), and
have a higher share of income from remittances than men.
The regression analysis based on the panel data and the 2013 cross-section yield results that do
not always tell the same story, probably reflecting the importance of controlling for
unobserved heterogeneity. The effect of nonfarm income on key household farming decisions
were examined in ten equations (counting the Two-Part models as two separate equations).
From the panel data estimates, a significant nonfarm participation (or nonfarm income) effect
is identified in only two of the ten equations—purchased input expenditure and total cultivated
area significantly increased with nonfarm income. The cross-section estimates tell the same
story with respect to purchased input expenditure but not cultivated area. Increased nonfarm
AFRINT 3: Ghana Micro Report
3
incomes were associated with reduction in cultivated area, suggesting competition rather
complementarity. The 2013 cross-section estimates also recorded significant effects of
nonfarm income on three other outcome variables of interest: hired labour use (positive
effect), rice output market participation (negative effect), and share of rice output sold
(positive effect).
Entering the gender decomposed nonfarm income variables into the regression equations, it is
observed that the probability of purchased input use and participation in nonfood cash crop
production are decreasing with male nonfarm income earnings while the share of rice output
sold is increasing with male nonfarm income earnings. Purchased input expenditure, the
probability of maize output market participation, and the probability of nonfood cash crop
production are all increasing with female nonfarm income; but total cultivated area is declining
with female participation in nonfarm income. Approximately 42% of households in the Afrint 3
sample have both male and female nonfarm income earners. We find that average purchased
input expenditure and the probability of improved seed adoption are higher among such
households; but such households also have smaller average cultivated areas.
Finally on welfare implications, it is first noted that composite welfare and food insecurity
status differs significantly across region and villages. Gyedi (in the Eastern region) has the
highest average value of the welfare index and the fewest number of households being food
insecure; Shia (in the Upper East region) is at the bottom. Welfare and food security highly
discriminates against living in the Upper East region compared with the Eastern region. We find
nonfarm incomes to be increasing across per capita income and wealth index quintiles,
suggesting that nonfarm income discriminates against the poor, although not in the sense of
participation because participation rates are not always increasing across the wealth
distribution. The regression results show that composite welfare increases with nonfarm
income but the magnitude of effect is often not of practical significance, particularly when
compared with other welfare enhancers such is human capital assets and livestock ownership.
While food insecurity is reducing with level of nonfarm income, there is a positive relationship
between nonfarm income participation food insecurity.
AFRINT 3: Ghana Micro Report
4
1. INTRODUCTION
Until relatively recently, much of the literature on farm-nonfarm-linkages have been largely
one-sided in that analyses of the link have been based on the supposition that it is growth in
the farm sector that precipitates events that lead to growth in the nonfarm sector (Mellor,
1976; Haggblade et al., 1989; Haggblade et al., 2007b; Diao et al., 2010). Even in this respect
foci have been on meso and macro level linkages. Although the role of nonfarm income in
triggering farm productivity and output has been conceived by a few during the time when the
above mentioned conventional wisdom thrived (e.g. Collier and Lal, 1984), the thinking that
nonfarm income could influence farm outcomes through its effect on farm investments arrived
later (see, for example, Evans and Ngau, 1991; Reardon et al., 1994; Savadogo et al., 1994;
Savadogo et al., 1995; de Janvry, 2005; Davis et al., 2009).
There are a number of reasons to expect nonfarm employment and income to influence farm
technology, production mix and farm outcomes in general. The adoption of new farming
technologies are potentially risky in sub-Saharan Africa (SSA) where the enabling environment
required for such technologies to thrive (water and other complementary input availability, for
example) are limiting. The availability of nonfarm income has the potential of mitigating such
risk and therefore likely increases the likelihood of technology adoption by smallholder farmers
(Evans and Ngau, 1991). If nonfarm incomes indeed reduce the tendency for self-provisioning
of household food requirements then, according to Reardon et al. (1994), one can expect the
availability of such income to influence a smallholder’s decision to increase cash crop farm size,
participate more in staple crop markets and sell a greater share of output, and use more
purchased inputs (including hired labour).
As noted in the next section, that nonfarm employment and income could have positive effects
on farm investments and outcomes are not established conceptually ex ante. For example,
empirical evidence from northern Burkina Faso in the late 1980s showed that households with
more nonfarm income invested less in farm capital (Christensen, 1989). As noted by Reardon et
al. (1994) investment of nonfarm incomes into farming depends chiefly on the preferred choice
of enterprise by the farm household in question as well as several conditioning factors
including agroclimatic conditions and infrastructure (both hard and soft), institutions (including
AFRINT 3: Ghana Micro Report
5
those central to the working of markets)2, type of nonfarm activity (i.e. sequential versus
contemporaneous), and who controls the nonfarm income in the household.3
This report contributes to the existing literature on the effect of nonfarm income on farm
activities and outcomes. In particular, the report extends existing knowledge by examining how
gender decomposed nonfarm participation influences farming activities and technology
adoption. The findings could be a source of hope or concern depending on the identified effect
of nonfarm income on farm capital and farm outcomes. If, for example, nonfarm income is
found to relax liquidity constraints so that participants are able to invest in improved farming
technologies that help raise output, then all else held constant, this could lead to agricultural
growth and poverty reduction. This should be the case if participation in nonfarm income does
not discriminate against the poor ab initio. On the other hand, if at the household level
nonfarm labour competes for farm labour in the presence of limited or absent hired labour
markets then farm output could suffer, and depending on the acquisition cost of farm output
forgone result in household food insecurity. In light of the possible welfare implications of
nonfarm participation the empirical work presented in this report encompasses the analysis of:
(i) the role of nonfarm income (including gendered intra-household nonfarm income) on farm
production technology and production mix; and (ii) the effect of nonfarm participation and
level of participation on household welfare and food security.
2. OVERVIEW OF THE LITERATURE
Interactions between farm and rural nonfarm employment follow four main interrelated
narratives: agricultural growth linkages (AGL), rural nonfarm employment (RNFE), household
livelihoods (HL), and regional development (RD) (Haggblade et al., 2007a). The first two are of
primary interest to the analysis in this report. The AGL narrative takes a sectoral perspective
and postulates synergies between the farm and rural nonfarm sectors (Hazell and Roell,
1983;Haggblade et al., 1989; Delgado et al., 1994). The distinctive feature of this model is its
focus on growth in the farm sector as the ‘engine’ that propels nonfarm activity and growth in
the rural economy (Hazell and Haggblede, 1993). The AGL model postulates employment
2 Where rural credit markets are limiting, for example, nonfarm income can be an important substitute. 3 In Ghana, for example, it will depend on existing rainfall patterns. For example, the northern parts of Ghana experience one rainfall season per year while the southern parts usually have a bimodal rainfall pattern. This has implications for both farm outcomes and participation in nonfarm employment.
AFRINT 3: Ghana Micro Report
6
linkages from the farm to the nonfarm sector (Reardon et al., 1998; Lanjouw and Lanjouw,
2001;Haggblade et al., 2002), production and consumption linkages (Haggblade and Hazell,
1989;Dorosh and Haggblade, 2003; Hossain, 2004; Anriquez and Daidone, 2010), and factor
market linkages (Reardon, 1997; Barrett et al., 2001a; de Janvry and Sadoulet, 2002;Foster and
Rosenzweig, 2004).
Rather than focus primarily on the farm sector as the engine the propels growth in the nonfarm
sector, and thus view farm-RNF linkages as a by-product of growth in the former, the RNFE
literature focuses on nonfarm employment in its own right. However, one of the main
conclusions of this narrative is that given rural household consumption preferences, rising
agricultural incomes will lead to higher expenditure on rural nonfarm output (Hazell et al.,
2007) leading to similar conclusions as the AGL model.
A catalogue of theoretically plausible effects of nonfarm activity on the farm sector can be
derived from the work of Ellis (2000). According to his analysis, the possible farm output effects
of nonfarm activities would depend on household labour allocation decisions, the relative
importance of agriculture in the future plans of a household as well as general social and
economic dynamics. Household asset endowment likely play an important role in the relative
importance of agriculture in the future plans of a household (Barrett et al., 2001b;Winters et
al., 2009) through inter-sectoral mobility effects. A brief overview of the empirical literature
that guided the analysis in this report is provided in the subsequent paragraphs.
Some have found complementarities between the farm and nonfarm sectors whereby capital
flows from nonfarm earnings to finance investment in agriculture at the household level (Evans
and Ngau, 1991;Reardon et al., 1992;Ellis and Freeman, 2004). In particular, wage employment
income can induce investment in farming only under restricted conditions of positive savings
and high nonfarm unemployment (Chikwama, 2004). On the other hand, it was observed in a
region of Ethiopia that increased farm output decreases participation in nonfarm wage
employment but increases participation in nonfarm self-employment in Ethiopia
(Woldenhanna and Oskam, 2001).
A series of articles in the Agricultural Economics journal of 2009 (vol. 40 no.2) focused on
household-level linkages between RNFE and farming. Most of the articles found positive effect
AFRINT 3: Ghana Micro Report
7
of RNFE on farm input demand and investment in agriculture. Two of the articles, Maertens
(2009) and Oseni and Winters (2009) , focused on Sub-Saharan African countries. Using a
sample of 240 households from a Senegalese region Maertens (ibid) found that households
involved in horticultural wage labour used greater quantities of purchased inputs and
cultivated their food crop farms more intensively. Oseni and Winters (ibid) analysed a
nationally representative rural household dataset on Nigeria and found farm input
expenditures to be increasing with RNFE activity. Similar results were found elsewhere (e.g.
Pfeiffer et al., 2009; Stampini and Davis, 2009; Takahashi and Otsuka, 2009). Hertz (2009), for
example, estimated a nonfarm income elasticity of purchased input expenditure of 0.14, an
estimate that is consistent with the farm credit constraint in the data used. In Albania, Kilic et
al. (2009) found that rather than investing nonfarm earnings into farming such income was
reinvested into facilitating movement away from farming.
There are a couple of relevant studies using data from Ghana. Canagarajah et al. (2001)
analysed rounds 1 and 3 of the nationally representative living standards survey data to
conclude that Ghana’s farm and nonfarm sectors were independent, that is, there were no
significant linkages. Per contra, Anriquez and Daidone (2010) employed the fourth round of the
survey to study linkages between the two sectors more explicitly and found significant cost
complementarities. So, the two studies tell different stories, suggesting change in conduct of
the rural economy of Ghana. But the differences in variables and methods employed could
account for the different conclusions reached even if identical data sets were employed. Other
researchers (Hilson, 2010; Okoh and Hilson, 2011) have used qualitative methods involving case
studies from mining areas of Ghana to show important synergies between artisanal and small-
scale mining activities and farming. So, even though the hypothesised link between nonfarm
activities and farm production is ex ante ambiguous (Ellis, 2000, p. 109), overall, household
level evidence suggests positive linkages between the two sectors through nonfarm income
effects on increased demand for purchased farm inputs.
AFRINT 3: Ghana Micro Report
8
3. DATA
The Ghana Afrint household surveys started in 2002 (known as Afrint 1) with a sample of 416
households drawn from eight villages located in the Upper East and Eastern Regions.4 The
primary focus then was on four food staples: maize, cassava, sorghum and rice. A second round
of surveys (Afrint 2) was undertaken in January 2008 with about 86% of households
successfully re-interviewed. Additional households were included in the Afrint 2 sample making
a total sample of 568 households. The 2008 survey instruments contained a large amount of
additional information. A major addition was questions on household income sources and
income.
In January 2013, a third round of data collection (Afrint 3) was conducted. A sample of 539
households was achieved during the survey. This is made up of 47 newly sampled households
and 492 or approximately 87% (including 3.7% descendant households) of the 568 Afrint 2
households (Table 1). A more thorough scrutiny of the sample, however, revealed that the
attrition rate was lower than it appears. The method used in drawing the initial sample (see
Dzanku and Sarpong, 2009) made it possible for two members of the same household to be
interviewed. There were at least 14 of such cases discovered in the Eastern Region,
predominantly in Gyedi. Taking this alone into account reduces the apparent attrition rate by
2.2 percentage points.
There were important modifications to the Afrint 3 household survey instrument. For example,
due to difficulties in obtaining reliable cassava production data (particularly output) focus on
this crop was drastically reduced. A major addition to the Afrint 3 household questionnaire was
the solicitation of gender disaggregated income data.
Given the focus on linkages between the farm and nonfarm sectors and the distributional
implications thereof a descriptive summary of the two sectors is first provided below. The
income descriptive analysis is based on the 2008 and 2013 datasets since the 2002 data does
not contain detailed income information.
4 See Dzanku and Sarpong (2009) for a more detailed description of the survey and sample. Note that the data reference year is always the year before the survey year.
AFRINT 3: Ghana Micro Report
9
(a) The rural farm sector
The rural farm households being studied are smallholders who cultivate an average of two
hectares (minimum and maximum cultivated areas are 0.04 and 10.5 hectares respectively) to
food and nonfood cash crops. We observe statistically significant differences in cultivated area
over time and region (Table 2). In both the Eastern and Upper East Regions average area
decreased between 2002 and 2008 (by approximately 11% and 5% respectively, on average)
before increasing over and above the 2002 size in 2013. Indeed, between 2008 and 2013
average total cultivated area more than doubled for the Eastern Region sample, the t-statistic
on this difference is 7.42 indicating high statistical significance. Given what we believe was a
better attempt at getting more accurate measures of cultivated area, which is a challenge in
most parts of Ghana, the large (and probably unrealistic) increase in farm size between 2008
and 2013 in the Eastern Region is probably a result of measurement error during the 2002 and
2008 surveys.
The Afrint 3 questionnaire contained the question: “If you compare your present farm size to
your farm size in 2008, has your farm size increased or decreased since then?” We compare the
responses to this question (i.e. area decreased since then, area unchanged or area increased
since then) with that calculated from reported cultivated area information. The column totals
(in brackets) show about 51% of farm managers reporting no change in cultivated area
between 2008 and 2013 (Table 3). Of the remainder households, 27% reported that their farm
sizes had increased with the rest indicating a decrease. Calculating the changes from total farm
size data reported by farm managers we observed an increase for 62% of all households (see
row total percentages in brackets) and a decrease for 33% of households, only about 5% of
households have unchanged size of cultivation area. Indeed, comparing the ‘reported’ and
‘calculated’ farm size changes we see that there is agreement for 32% of cases, indicating the
possibility of measurement error in a number of cases.
Turning again to Table 2 we observe that aside from maize, cassava, sorghum and rice, most
farm households in our sample cultivate other food crops. The most popular among the list in
the Eastern Region in 2008 were cocoyam (78% participation), plantain (74% participation),
vegetables of local markets (70% participation) and yam (62% participation). Exactly the same
order was maintained in 2013 but participation increased for cocoyam (82%) and plantain
(81%), decreased for vegetables (68%) and remained largely unchanged for yam. For the Upper
AFRINT 3: Ghana Micro Report
10
East Region the most widely grown crops in 2008 were: groundnuts (91%), beans (80%), millet
(73%) and vegetables (67%). The order changed in 2013 as follows: millet (98%), groundnuts
(88%), beans (73%) and vegetables (70%).
Aside other food crops, some households also grow nonfood cash crops. We observe (Table 2)
that while participation in nonfood cash crop production increased systematically between
2002 and 2013 in the Eastern Region (from 10.6% of households in 2002 to 42.1% in 2013),
participation declined consistently in the Upper East (from 11.1% in 2002 to 1.8% in 2013).
Anecdotal evidence from our interactions in the villages suggest that growing preference for
maize which serves as both cash and food crop is partly responsible for this decline. The main
nonfood cash crops were cocoa and oil palm in the Eastern Region, and tobacco in the Upper
East. For households that cultivate nonfood cash crops, a larger average area is devoted to such
crops than to ‘other’ food crops (Table 2). We now turn attention to the three staple crops
studied in detail: maize, sorghum and rice.
Farm size, output and yields
The study collected farm size and output information covering the immediate three seasons
prior to the surveys. This makes available a total of nine data points on farm size and
production (Table 4). Participation in maize production during the Afrint 1 period (the 1999
production season through 2001) reached nearly 100% in the Eastern Region where maize is
the most important staple food crop, but declined to 94% during the Afrint 2 period (from 2005
through 2007), before increasing to about 97% during the most recent survey (2010-2012).
Maize production information was not collected for the Upper East Region during the Afrint 1
period. For the Afrint 2 & 3 periods we observe, as expected, lower participation in maize
production in the Upper East than in the Eastern Region. However, participation has been
increasing in the Upper East (from about 32% of households in 2005 to 62% in 2012). Focused
group discussions suggest that maize is gradually replacing sorghum in the Upper East villages.
Not surprising, we observe a significantly reduced participation in sorghum production during
the Afrint 3 period (86% participation compared with 94% during the most recent Afrint 2
season). Participation in rice production decreased during the Afrint 2 period compared with
Afrint 1 but increased during Afrint 3 although still below the Afrint 1 participation rate (Table
4).
AFRINT 3: Ghana Micro Report
11
Maize farm size decreased consistently during the Afrint 1 & 2 periods, starting with mean size
of just about a hectare in 1999 and reducing to 0.64 ha by the 2007 growing season (Table 5).
The Afrint 3 period, beginning at the 2010 growing season, recorded a large increase in farm
size compared with the Afrint 2 period. Mean maize farm size was 1.12 ha in 2010 and 1.14 ha
in 2012. Table 6 shows a statistically significant 31% decrease in mean maize farm size in the
Eastern Region between Afrint 1 and 2. However, mean maize farm size for Afrint 3 is
significantly greater than that for Afrint 1 and 2 by 22% and 77% respectively. Average maize
farm size did not differ significantly across region during the Afrint 2 periods. Within the Upper
East, average maize farm size remained largely unchanged between the Afrint 2 and 3 periods
(Table 6). Average sorghum farm size increased significantly between Afrint 1 and 2 as well as
between Afrint 1 and 3; between Afrint 2 and 3, however, there was no significant change.
Mean rice farm size remained largely constant between Afrint 1 and 2 but increased
significantly between Afrint 1 and 3 and between Afrint 2 and 3 (Tables 5 and 6).
We observe that as maize farm size declined between Afrint 1 and 2, output also declined—
households were producing an average of 790 kg – 890 kg of maize in the Eastern Region
during Afrint 1 but this decreased to 635 kg – 765 kg during the Afrint 2 period (Table 5). As
average farm size increased during Afrint 3 output also increased (from 995 kg – 1,145 kg).
Because the change in output was offset by the change in farm size between the periods we
observe that, at the 5% level of significance, there was no change in average maize yields over
the three surveys in the Eastern Region (Table 6). In the Upper East Regions, however, because
the increase in maize output outpaced the increase in average farm size we observe that
average maize yield grew by more than 200% and is statistically significant.
An average household was producing 382 kg of sorghum and 496 kg of rice during Afrint 1;
sorghum output dropped to only 141 kg and rice to 309 kg in Afrint 2 before rising to 258 kg in
the case of sorghum and 684 kg in the case of rice during Afrint 3. These production figures
meant that both sorghum and rice yields declined between Afrint 1 and 2 and between Afrint 1
and 3. Yields recorded in Afrint 3, however, represent a 167% and a 98% increase over levels in
Afrint 2 for sorghum and rice respectively.
Input use
AFRINT 3: Ghana Micro Report
12
We provide a snapshot of farm input use among households over time and across region in
Tables 7 and 8. During the 2008 survey we observed a statistically significant increase in the
proportion of maize farmers using improved seeds in the Eastern Region, from approximately
37% in 2002 to 64% in 2008. The recent survey shows a reduction in the proportion of farmers
using improved maize seed, down to 56%, but the reduction is not statistically significant at the
5% level. For the Upper East Region sample, we observe a drastic reduction in the proportion of
maize farmers using improved seed, from about 65% of farmers in 2008 to only about 2% in
2013. The reason for this reduction is that during the Afrint 2 period the Ministry of Food and
Agriculture was actively supplying improved seeds in the Upper East Region villages or the
environs but this had waned during Afrint 3 although seeds were available in the market. The
situation was similar for improved rice seed use: only 8% of farmers were using improved seeds
in 2013 compared with 32% in 2002 and 64% in 2008. As for sorghum all farmers use traditional
varieties.
Although still less than half of maize farmers were using inorganic fertilizers in the Eastern
region, the proportion using the input has been increasing consistently over the survey years,
from 24% in 2002 to 35% in 2008 and then to 42% in 2013 (Table 7). Real average fertilizer
expenditures (after adjusting for general price level changes) for those using the input
decreased from about US$28 in 2008 to US$23 in 2013, the t-statistic on this difference is 1.75
indicating lack of statistical significance at the 5% level. In the Upper East Region, a higher
proportion of maize farmers (70%) were using fertilizers in 2013, up from about 45% in 2008
(Table 7). However, real expenditures on the input decreased from approximately US$28 in
2008 to US$20 in 2013, the difference is significant at the 1% level. Although there was an on-
going national fertilizer subsidy programme no farmer reported participation in the
programme.
Only about 14% of sorghum producers were using fertilizers in 2002 but this dropped further to
just 4% of farmers in 2008 before increasing to 16% in 2013 (Table 7). Those using fertilizer on
sorghum were spending an average of US$16 on the input in 2008 but this declined to US$12 in
2013, but this difference is due to chance variation (t-statistic = 1.2) . A higher proportion of
rice farmers (than sorghum farmers) used fertilisers and also spent more real US dollars on the
input. The t-statistic on the decrease in mean real rice fertilizer expenditures between 2008
and 2013 is 0.83, indicating no statistically significant decrease over the period.
AFRINT 3: Ghana Micro Report
13
Pesticide and herbicide use on maize, sorghum and rice across regions increased consistently
over the three surveys (Table 7). Herbicide use as a land preparation method has become very
common in the Eastern Region study villages where focus group discussions revealed that this
is a result of the relatively expensive cost of hired labour for land preparation. However, we do
not observe a significant reduction in hired labour use from the household data over the panel.
Indeed, hiring labour for farm activities is more common in the Eastern than the Upper East
Region (Table 8).
We show changes in household use/access to other inputs in Table 8. These include the use of
animal manure, contact with government agricultural extension agents, and access to input
credit. Animal manure is mainly used in the Upper East Region with 81% of farmers using it in
2002, 92% in 2008 and 88% in 2013. The proportion of farmers reporting contact with
agricultural extension agents increased from 57% to 64% between 2002 and 2008 in the
Eastern region but decreased from 84% to 50% in the Upper East over the same period. By
2013, agricultural extension agent contact decreased among farmers in both regions, down to
52% and 49% in the Eastern and Upper East respectively. Statistically, only the decrease
between 2008 and 2013 was significant for the Eastern Region while for the Upper East only
the marginal decrease between 2008 and 2013 was not significant. Except in 2013, the
proportion of farmers reporting agriculture extension contact has been significantly higher in
the Upper East than the Eastern Region.
Farmer organization membership has significantly decreased in the Upper East from one survey
to the other; in the Eastern region the decrease between the 2008 and 2013 surveys was not
significantly different from zero (Table 8). Agricultural input credit is not common in the study
villages; for all the three surveys less than 15% of farm households reported such credit, which
in most cases comes from private individuals, often farm produce aggregators (Table 8).
Output marketing
It has been noted (see, for example, Barrett, 2008) that staple crops are not sold by a large
proportion of rural farm households, and that agroecological potential is an important
determinant of staple crop output sale decisions. We see from Table 9 that very few
households in the Eastern region (about 4%) did not sell maize during the first two surveys. The
proportion that sold maize during the most recent survey decreased such that nearly 16%
AFRINT 3: Ghana Micro Report
14
reported no maize sales. The high participation is because, although maize is the single most
import cereal staple crop in Ghana (Angelucci, 2012) it is also considered an important cash
crop. The effect of agroecological potential differences on market participation is clearly visible
as we observe that in 2008 only about 3% of maize producing households in the Upper East
Region sold some of the output. By 2013, the proportion selling increased to nearly 22%. This is
not surprising for two reasons. First, flooding in the Upper East Region during the 2007 crop
year affected production in some of the study villages. Second, as noted earlier, maize is
gradually replacing sorghum in the Upper East study villages due to declining yields.
Households selling maize put an average of between 60% (in 2002) to 66% (in 2008) of their
output on the market in the Eastern Region compared with about 38%-42% of output in the
Upper East Region (Table 9). In terms of actual sale quantities for the two equivalent survey
years (i.e. 2008 and 2013) the pooled average was 642 kg for the Eastern Region and 287 kg for
the Upper East, the difference being highly significant.
Sorghum is produced mainly for home consumption in the Upper East villages but one out of
every four households was selling an average of one-third of output in 2002. As indicated
earlier the precarious conditions faced by some households in 2008 meant that only about 3%
of households reported some sorghum sales in 2008. But even in a ‘normal’ year (i.e. 2012)
only a little over 10% of sorghum producers put some of their output on the market in 2013.
Rice serves as both a staple and cash crop in the Upper East so we see that 56% of rice farmers
were selling (an average of 332 kg) in 2002 and even under harsh climatic conditions in 2008
32% put some rice (mean of 246 kg) on the market. In 2013 51% of rice producers sold an
average of 725 kg of rice representing 52% of their mean output.
(b) The rural nonfarm sector
Rural nonfarm employment is an important part of rural household livelihoods (e.g. Ellis, 2000;
Lanjouw and Lanjouw, 2001; Haggblade et al., 2007b; Davis et al., 2010; Ellis, 2010). The
descriptive analysis of the rural nonfarm sector provided here is based on the Afrint 2 and 3
surveys. In these surveys information was solicited concerning 12 income sources: sale of food
staples, sale of other food crops, sale of non-food cash crops, sale of animals/animal produce,
leasing out machinery and/or equipment, work on others’ farms/agricultural labour, non-farm
salaried employment, micro business, large-scale business, rent and interest, pensions, and
AFRINT 3: Ghana Micro Report
15
remittances. Considering these income source we observe that the average Eastern Region
household had 2.8 (minimum of 1 and maximum of 7) income sources in 2008 and 3.3 in 2013
(also ranging between 1 and 7), a statistically significant increase of 20% (Table 10). For Upper
East households the average number of income sources rose from 2.4 in 2008 to 3.2 in 2013
representing a significant increase of 31%. The regional difference in average number of
income sources per household is significant for the 2008 sample but not 2013.
Household average incomes (both current and real) were significantly higher in the Eastern
than Upper East Region in both 2008 and 2013, as one might expect. Over the two periods,
there was no significant change in real average household income in the Eastern region, but
real income rose by 39% in the Upper East Regions (Table 10). After accounting for household
size, however, we observe no significant change in income over time in both regions; there
were, however, significant differences in per capita income by region.
For a somewhat more meaningful and concise descriptive analysis we put the income sources
into seven groups: food crops, nonfood cash crops, livestock, non-labour, nonfarm wage
employment, nonfarm self-employment, and remittances (Table 11). First, at a higher level of
aggregation, we observe that income from crops accounted for the largest share of average
household income in the Eastern Region (77% in 2008 and 64% in 2013). Crop income was far
less important in the Upper East where in 2008 only about 19% of average total income was
from crops, increasing to 29% in 2013. These changes over time within region as well as across
region are significant at the 1% level (Table 12). While there was near perfect participation in
crop income in the Eastern Region in both survey years (98% in 2008 and 95% in 2013), only
39% and 62% of Upper East Region households received some income from crops in 2008 and
2013, respectively. Even after conditioning on participation, still less than half of average Upper
East household income came from crops.
Nonfarm income as a whole (income from sources other than crops and livestock but including
working on other peoples’ farms) accounted for 16% and 43% of average total income in the
Eastern and Upper East Regions respectively in 2008. By 2013 average nonfarm income shares
had increased to 32% in the Eastern Region and 51% in the Upper East Region. These changes
and differences are strongly significant at conventional levels of testing (Table 12). Clearly,
AFRINT 3: Ghana Micro Report
16
nonfarm income is relatively more important in the poorer and lower agroecological potential
region with a monomodal rainfall pattern which gives rise to a long nonfarm season.
In the two regions, nearly all of the crop income came from food crops, not nonfood cash
crops—only 2.5% and 5.3% of average income in the Eastern Region and 1.1% and 0.5% in the
Upper East in 2008 and 2013, respectively, came from nonfood cash crops. Participation in
nonfood cash crop income is low in general but very low (4.5% in 2008 and 1.5% in 2013) in the
Upper East Region. Participation is much higher in the Eastern Region, up from 11% of
households in 2008 to 29% in 2013. Even households participating in nonfood cash crop
income, on average, receive a greater share of income from food crops.
Income from livestock completes the components of farm income and is clearly relatively more
important in the Upper East than the Eastern region. Approximately 6% and 4% of average
income was from livestock in the Eastern Region compared with 38% and 20% for the Upper
East, over the two periods. Also, a larger share of households in the Upper East (80% in 2008
and 74% in 2013) than in the Eastern Region (42% in 2008 and 37% in 2013) obtained income
from livestock.
The relative importance of nonfarm income sources differ by region. On average, nonfarm self-
employment income is the most important in terms of contribution to household income in the
Eastern Region, but contributed only 7% of average total household income in 2008 and
approximately 17% in 2013. For the Upper East Region, remittance inflows from absent family
members contributed the most to household income—approximately 17% in both years—than
any other nonfarm income source (Table 11). Nonfarm wage employment and remittances
were the second and third most important nonfarm income source in the Eastern Region in
2008 but accounted for less than 5% of total income in each case. In 2013, remittances become
more important that nonfarm wage employment in the Eastern region. As for the Upper East
Region, beside remittances, nonfarm wage employment was more important than self-
employment, both in terms of participation and income shares, for both 2008 and 2013.
The 2013 survey collected gender disaggregated income data which we explore in Tables 13
and 14. First, in Table 13 we compare income and income source variables by household
headship as well as intra-household gender differences. The 2013 survey covered 534
AFRINT 3: Ghana Micro Report
17
households who reported information on their incomes and income sources. Approximately
20% of the 534 households were female headed. At the 5% level, we observe statistically
significant differences between the two types of households in 8 out of 12 variables reported in
Table 14. The exceptions are per capita income, crop income share, nonfarm self-employment
income and its share in total income. Female headed households have a smaller number of
income sources, a smaller number of members in nonfarm work, higher crop income, more
income from remittances, higher remittance income share, lower nonfarm income, but higher
nonfarm income share.
Moving to the intra-household descriptive analysis, first, we observe that out of the 534
households, 59% have both male and female income earners. In total, however, there are at
least 851 individual income earners in the surveyed households, of which 47% are females. Out
of the 11 relevant variables in Table 13, there exist significant intra-household gender gaps in 9.
The two exceptions were nonfarm self-employment income and remittances for the unpaired
mean differences; and remittances and nonfarm income for the paired comparison. Otherwise,
females have statistically significant fewer income sources, smaller income earnings, smaller
crop incomes and crop income shares for both the paired and unpaired analysis, and in
addition have smaller nonfarm incomes in the case of the unpaired means.
Within households, more females than males are involved in nonfarm work, have higher share
of income from nonfarm self-employment and remittances, and indeed have higher share of
their incomes from nonfarm income in general. For households with both male and female
income earners, females earn higher average income from nonfarm self-employment—
approximately US$509 compared with US$267 for males (Table 13).
We provide detail gender and regional disaggregated income shares and participation
descriptive statistics in Table 14. In most cases we observe positive and often significant gender
gaps (meaning higher income shares or participation for males), particularly when not
conditioned on participation. The three important exceptions where we observed negative
intra-household gander gaps for both regions were mean nonfarm self-employment income
shares, remittance share and overall nonfarm income share. For example, in the Eastern Region
an average of about 53% of female income came from nonfarm sources with average
participation rate of 80% compared with nearly 27% nonfarm income share for males with 58%
AFRINT 3: Ghana Micro Report
18
participation. The importance of nonfarm income for females is even more marked in the
Upper East Region where, on average, 74% of all female incomes were generated in the
nonfarm sector with participation rate of 94%. Compare this with male nonfarm income share
of 48% and participation rate of 88%.
So, overall, how much do women contribute directly to household average cash income in our
sample? In the entire sample at the village level this ranges from 27% in Apaa (in the Eastern
Region) to 44% in Shia (Upper East region)—the regional averages are 37% and 36% in the
Eastern and Upper East regions respectively, the difference being statistically insignificant.
Considering only households with both male and female income earners, average female
contribution ranges between 28% of household income in Apaa to 48% in Asitey; the regional
average is 36%, meaning that females were contributing about 36% of total household income.
4. ANALYSING FARM–NONFARM LINKAGES AND DISTRIBUTIONAL ISSUES
The main thrust of this paper is the analysis of the household-level farm-nonfarm linkages and
the distributional implications thereof. Linkages between the farm and the nonfarm sectors
have been the focus of past research, and in recent times some attention has been given to
household level linkages as shown in section 2. The analysis here draws largely from this
literature, mutatis mutandis, as it adds on the distributional aspects.
(a) Analytical methods
We rely on descriptive and regression analysis for gaging the link between household nonfarm
activities and farm production behaviour as well as the distributional implications. The
descriptive analysis uses mainly bivariate analytical tools. We specify the regression models
below.
First of all, we are interested in the effect of participation in nonfarm employment or earnings
on farm outcomes (i.e. farm output and productivity). Since it is hypothesised that nonfarm
earnings likely reduces liquidity constraints, particularly in the presence of credit market
failures, it makes sense to conceive that nonfarm participation or earnings affect output and
productivity through their effect on production input use decisions. With this in mind we
specify the following general regression equation:
AFRINT 3: Ghana Micro Report
19
,it it it i itY nfm X cα β δ ε= + ∗ + ∗ + + (1)
where i and t indexes household and time respectively; Y is the farm related outcome of
interest (e.g. purchased input use, hired labour, area under cultivation among others discussed
below); α is the intercept term; β is the coefficient on the nonfarm participation or earning
variable nfm, and is of primary interest; X is a vector of individual, household and farm
characteristics; ci is the household specific effects or heterogeneity assumed to be time-
invariant; and ε is the idiosyncratic error term.
Clearly, an estimate of the marginal effect of nfm on Y, β, is biased if unobserved individual and
household characteristics that influence participation in nonfarm activities also affect the
outcome decisions of interest. Allowing household specific heterogeneity ci to be correlated
with nfm and X could take care of this identification problem, making β an unbiased estimate of
the marginal effect of interest. Proceeding this way assumes that endogeneity operates
through omitted heterogeneity only. However, it is possible that endogeneity arises through
correlation between nfm and ε as well. In this case we consider estimating the structural and
reduced for equations simultaneously or use other methods (described below) to account for
such possible endogeneity.
We estimate the effect of nonfarm participation/earnings on farm outcomes using six
dependent variables: expenditure on purchased inputs (fertilizer, herbicides and pesticides),
improved seed adoption, hired labour use, staple crop output market participation,
participation in nonfood cash crop production, and total cultivated area.
Aside from area under cultivation, all the other dependent variables are characterised by a
fairly large mass at zero. There are at least three approaches to modelling such variables: Tobit
(Tobin, 1958), selection models based on the seminal article by Heckman (1979), and two-part
models (Cragg, 1971; Duan et al., 1984).
In the present case, the zeros are actually observed data. This is because the population of
interest from which the data was taken are all agricultural producers and not a self-selected
sample. Thus, theoretically, one is not faced with a sample selection problem per se (see Hertz,
2009 for a similar argument). The Tobit model is the most frequently applied in related
literature (e.g. Kilic et al., 2009; Pfeiffer et al., 2009; Takahashi and Otsuka, 2009). However, it
AFRINT 3: Ghana Micro Report
20
is important to note its shortcomings: reliance on homoscedasticity and normality of the error
term for consistency as well as the assumption that the process generating the zeros and
positive outcomes are essentially the same. The two-part model (TPM) estimates the
probability of a positive outcome in the first part, and the magnitude of the positive outcome in
the second part. This is the preferred choice in the analyses, and Tobit models are used only for
comparison.
The general form of the relevant Tobit model can be written as:
*
* *
,
if 0,0 otherwise
it it it i it
it itit
Y nfm X c
Y YY
α β δ ε= + ∗ + ∗ + +
>=
(2)
where Yit is the observed dependent variable and *itY is the latent variable which is related to
the observed as stated above; all other variables are as described in equation (1). Estimating
equation (2) in the presence of nfm being potentially endogenous presents an econometric
challenge because nfm itself is semi-continuous (i.e. a substantial number of households
neither work nonfarm nor receive nonfarm income).
This challenge can, in part, be surmounted by writing the equations with a bivariate Tobit
model structure assuming the error terms are bivariate normally distributed (Amemiya, 1974).
*1 1 1 1 1
*2 2 2 2 2
it it it i it
it it i it
Y nfm X c
nfm X c
α β δ ε
α δ ε
= + + + +
= + + + (3)
where it is assumed that 2 21 2 1 2 12( , ) ~ (0,0, , , ).it it Normalε ε σ σ σ This deals with endogeneity
arising particularly from simultaneity bias (Chen and Zhou, 2011). But then one has to find a
way of sweeping out the unobserved heterogeneity. Given that the time dimension of the
panel is the minimum possible (i.e. T = 2), the work of William Greene (see Greene, 2004b;
Greene, 2004a) suggests that pooling the data is not a bad idea in the presence of the
incidental parameter.
For the two-part model (TPM) we specify a probit model for the first part as:
AFRINT 3: Ghana Micro Report
21
*
*1[ 0]~ (0,1).
it it it i it
it it
it
Y nfm X c
Y YNormal
α β δ ε
ε
= + + + +
= > (4)
In order to deal with potential endogeneity of nfm we treat it us a dummy variable, which in
itself has important advantages. Aside from making it possible to estimate the average
treatment effect of participation in nonfarm employment or income, it reduces the chances of
measurement error which could be a big problem with household income data (Deaton, 1997;
Stampini and Davis, 2009). With this in mind a bivariate probit model is applicable in this
instance. Another estimation option for the first part is a linear probability model (Angrist,
2001). For panel data, a fixed effects linear probability model (e.g.Bandiera, 2007; Deininger
and Ali, 2008) is useful for modelling unobserved heterogeneity.
The second part of the TPM (i.e. for Yit > 0) is:
.it it it i itLog Y nfm X cα β δ ε= + + + + (5)
Again, one has to deal with the possibility that E(ε|nfm) ≠ 0. If we allow nfm to enter as a
dummy, then nfm can be seen as capturing the average treatment effect so that equation (5)
can be estimated under the treatment-effect model framework where we use the 2013 data
only or pool the data.5 Where nfm enters as level of nonfarm income and is treated as
endogenous, correction terms are generated from a first-step pooled Tobit or random effects
Tobit models and added as additional regressors in the farm outcome equation of interest
(Vella, 1993; Vella and Verbeek, 1999).
There is one more estimation issue to address, which is in the case where the share of staples
sold is the dependent variable. This is a fractional response variable, 0 ≤ yit ≤ 1, with outcomes
at the endpoints, zero and one inclusive. In this case, applying the approaches described above
could be inappropriate because they cannot ensure that the predicted values of the response
variables, given the entire continuous distribution of explanatory variables, lie within the
interval of the bounded dependent variable (Papke and Wooldridge, 1996, 2008). Under strict
exogeneity the model is specified in a general form as: 5 We estimate a number of models using the panel data and the 2013 data only because the later allows us to test additional hypotheses which are not possible using the panel. This is because data on some variables were not collected in 2008.
AFRINT 3: Ghana Micro Report
22
( | , ) ( )it i i it it iE y X c G nfm X cα β δ= + + + (6)
where G(∙) is either a logistic or normal distribution function. We work with G(∙) ≡ Ф(∙) and
apply the Bernoulli quasi-maximum likelihood estimator (QMLE) following (Papke and
Wooldridge, 2008). The panel data component can also apply the generalised estimating
equation approach (Zeger and Liang, 1986), allowing misspecification of the model error
structure (Papke and Wooldridge, 2008).
Next the study seeks to assess the welfare effects of participation in nonfarm work or income.
Specifically, interest is in the effect of nonfarm income on household welfare. Two indicators of
welfare are used: a composite welfare index and food security. The welfare index equation is:
,it it it i itW nfm X c uγ λ η= + + + + (7)
where W is the fully observed welfare index described below;γ is the intercept term;λ is the
unknown parameter of primary interest;η is the a vector of unknown parameters associated
with the vector X containing exogenous household and individual characteristics, assets,
location dummies; and uit is the error term. Treating nfm as exogenous in equation (7) may
lead to the estimate ofλ being biased. This is because households or individuals may choose to
participate in nonfarm work or not conditional on unobserved characteristics. Indeed, better-
off households may choose to participate in high-return nonfarm activities, and entry barriers
may exclude the poor (Barrett et al., 2001a). Similarly, for low-return type nonfarm work, the
relatively wealthy may choose not to participate.
In the panel data context we sweep out the unobserved effects that may cause the bias using
the fixed effects estimator (which is equivalent to first-differencing for T = 2). Supposing this
does not suffice we generate correction terms from a pooled or random effects Tobit model for
level of nonfarm income to be included in the welfare equation as additional regressors to
correct for any bias that may still be working through the idiosyncratic error (Vella, 1993; Vella
and Verbeek, 1999).
Where we use the 2013 cross-sectional data only or pool the data, equation (7) is simply an
endogenous dummy-variable model which is estimable under the endogenous treatment-
regression framework:
AFRINT 3: Ghana Micro Report
23
1, if 0
0, if otherwise
j j j j
j jj
y X nfm
L eoffm
β π n
t
= + +
+ >=
(8)
where Xj is the vector of explanatory variables in the welfare equation; Lj is a vector of
covariates that explain participation in nonfarm income; and vj and ej are the error terms
assumed to be bivariate normally distributed with zero mean and covariance matrix
2 1
σ ρσρσ
.
Finally, we estimate the food security effect of nonfarm participation. Since the food security
indicator is binary we allow the nonfarm participation variable to also enter the model as such
to make for a simpler estimation procedure when having to deal with endogeneity (Greene,
2012). In this case we specify a general bivariate probit model as
* *1 1 1
* *2 2 2
11 2
2
, 1 if 0, 0 otherwise,
, 1 if 0,0 otherwise,
0 1 | , ~ , .
0 1
fs X nfm v fs fsnfm X v nfm nfm
vX X Normal
v
β η
β
ρρ
′= + + = >
′= + = >
(9)
Where panel data is used, (9) a correlated random effect bivariate probit model is used where
correlation between the unobserved effect and the covariates is captured by group mean
variable addition.
5. RESULTS
Results of the regression analysis are presented here. Prior to that, descriptive analyses of the
main issues appear first.
AFRINT 3: Ghana Micro Report
24
(a) Descriptive analysis
Most of the variables that appear in the regression analyses have already been discussed. Table
18 presents a summary of all variables by nonfarm income participation status. In the Afrint 3
dataset, only 13% of all households did not report any nonfarm incomes.6 In the pooled panel
data, the proportion of households not reporting nonfarm income is 31% (42% during Afrint 2
and only 18% during Afrint 3).7 We focus on a few selected variables in Table 18. Mean per
capita cash income of Afrint 3 households was higher among non-participants than
participants; the average difference of $95 is significant at conventional levels. Using the
detailed gender-disaggregated income data, the average income obtained by non-participant
households was about $200 higher than for participants. This suggests that receiving nonfarm
income was not associated with higher overall incomes, on average.
It is observed that average female income as percentage of average household income is
statistically lower among non-participants than participants by approximately 12%. This is not
surprising since female members tend to be more involved in nonfarm income generating
activities than males. For example, there were about 11% more female-headed households
among nonfarm income participants than non-participants.
Nonfarm income is hypothesised to improve farm productivity through its potential effect on
farm inputs (including hired labour). As Table 18 shows, the proportion of households using
purchased inputs is significantly higher by approximately 14% among non-participants than
participants during Afrint 3, a result contrary to expectation. However, the extent of use does
not differ across the two groups. Also, more non-participants than participants (difference of
28% during Afrint 3 and 18% in the pooled sample) were using improved seeds. The use of
hired labour does not differ significantly across the two groups.
One would expect both food crop market participation and nonfood cash crop production to
increase with participation in nonfarm income because the latter is expected to serve as a
buffer that reduces the orientation towards ‘safety first food cropping’ and subsistence
behaviour (Reardon et al., 1994). There is contrary evidence from the descriptives: a
significantly larger proportion of nonfarm income non-participants than participants sell own- 6 The proportion not reporting nonfarm incomes increases when not using the detailed gender disaggregated income data, up by approximately seven percentage points. 7 Note that the pooled data does not contain the refreshment sample.
AFRINT 3: Ghana Micro Report
25
produced maize and rice; also the share of maize output sold by non-participant is significantly
larger than that sold by participants. It could be that households participating in nonfarm
income generating activities produce staples mainly for consumption while non-participants
rely more on this crops for both consumption and income. Aside food crops, it is also observed
that a lower proportion of nonfarm income participants participate in commercial vegetable
and non-food cash crop production, the difference is significant at conventional levels (Table
18).
A major reason farmers give for their inability to expand their cultivated area is the lack of
liquidity. If participation in nonfarm income reduces this liquidity constraint then one would
expect that participants in nonfarm income would, on average, cultivate larger areas if there
are complementarities between the two sectors. Otherwise, the nonfarm sector may be seen
as competing with the farm sector, barring the contribution of the former to intensification in
the latter. In the pooled panel data, no significant difference in cultivation area is observed
between the two groups. In the Afrint 3 sample, however, non-participants cultivated 0.68 ha
more land than participants, and this difference is significant.
Turning to the descriptives on the distributional implications of nonfarm income participation,
two measures of welfare are used: a composite welfare indicator and indicators of food
(in)security. First, some comments on Table 19 which contains the descriptives on the welfare
indicators. The welfare indicator was constructed by aggregating ownership of household
durables (mobile phone, motor bike, televisions, sowing machine, sofa set), a household’s
ability to save money, household dwelling characteristics, and household non-labour income
(see, for example, Finan et al., 2005).
On food in(security), a household is defined as food insecure in the panel dataset if the number
of meals eaten per day during the lean season was less than that eaten during the rest of the
year. During Afrint 3, additional questions allows construction of a food (in)security indicator
based on three dimensions: meal quantity, quality, and frequency. A binary food insecurity
indicator is defined that takes on the value one if a household reduces meal quantity, quality
and frequency during the lean season compared to the rest of the year. Additionally, a food
security index is constructed which takes on the values zero through three: zero means
reduction in all three dimensions; one means reduction in any two dimensions; two means
AFRINT 3: Ghana Micro Report
26
reduction in any one dimension; and three means meal quantity, quality, and frequency
remains constant throughout the year.
Table 19 presents a summary of the welfare indicators by household location. The village with
the highest value of the average welfare index is Gyedi and the lowest in Shia. Statistical tests
(t-tests from a regression of welfare on village dummies only) show that the average household
in all other villages has significantly lower welfare than the average household in Gyedi.
Normalising using Gyedi’s average value of the index, Shia, for example, has only 43% of the
average welfare value of households in Gyedi. The village with the second highest average
value of the index, Asitey, has 77% of Gyedi’s average. The panel data shows average welfare
increased over time in all villages, averaging a 69% increase across all villages over the panel
period (or about 14% per annum).
It is noted using ofs4 (see Table 19) that, overall, the village with the lowest and highest values
of the welfare index are also those that have the lowest and highest proportion of households
being food secure. For example, in Gyedi 82% of households maintain the quantity, quality and
frequency of meals all year round, but in Shia only 11% do same. Statistical tests (z-statistics
from a probit regression of ofs4 on village dummies only) show that the likelihood of being
food-secure is statistically lower in all other villages than Gyedi—the exceptions are Asitey and
Apaa, all in the Eastern Region.
Returning to Table 18, it is expected, crudely, that if participation in nonfarm income has
positive distributional implications then participants should have, on average, higher values of
the welfare index and have higher proportion of food-secure households. This is the case for
the welfare index in the pooled panel data set but the contrary is observed in the Afrint 3 data.
As for food security, both the panel data and the Afrint 3 data tell a consistent story but not
according to expectation: a larger proportion of non-participants in nonfarm income are food-
secure compared to participants, the difference of 12% and 13% in the pooled panel and the
Afrint 3 data are both significant at conventional levels of testing.
A bit more detail on the distributional implication can be found in Table 20. Households are
grouped by per capita income and welfare index quintiles to help understand the distribution
of household income, nonfarm income, income shares and their disaggregation by gender
AFRINT 3: Ghana Micro Report
27
across groups. The gap in average total household cash income between the lowest income
quintile group and the highest is wide: approximately 26 times for the Afrint 3 sample and even
wider in the pooled panel data. Although per capita income is also increasing consistently
across welfare index quintiles, the gap is not so wide. For example, the highest welfare quintile
households received only 4.5 and 5.3 times more income than the lowest group in the Afrint 3
and pooled panel, respectively.
It can also be seen in Table 20 that nonfarm incomes are increasing as we move from the
lowest to the highest per capita income and wealth index quintiles, which may be viewed as
prima facie evidence that nonfarm incomes discriminate against the poor. But, in fact, Table 20
shows that the poor are not participating any less than the rich in nonfarm incomes—nonfarm
income participation rates are mostly not increasing throughout the per capita income and
welfare index quintiles. A little more critical scrutiny of the data shows that the poor are
deriving about the same or higher shares of their income from nonfarm sources, suggesting
that they are likely involved in low-return type nonfarm activities (Table 20).
Finally, we see that both male and female nonfarm incomes are increasing with the
constructed income and welfare index quintiles, but the gaps appear wider with male than
female incomes. For example, average male income at the highest wealth index quintile is six
times that at the lowest quintile while for female incomes the gap is only three. Female
nonfarm income shares are also highest among the poorest households. For instance, female
average nonfarm income share is about 54% among per capita income poorest households
compared to 35% among the richest. This would suggest that poorer households have females
contributing more of nonfarm incomes than rich households. The story is not different with
respect to women’s share of overall household income. These findings are generally consistent
with those of Reardon (1997) and Owusu et al. (2011), for example, the later using data from
villages in the Northern region of Ghana.
(b) Regression results
The focus of the regressions is twofold: to analyse the effect of nonfarm income on farm
outcomes; and to examine the welfare implications of participation in nonfarm income. Prior to
discussing the results addressing these, the determinants of nonfarm income participation and
extent are briefly explored.
AFRINT 3: Ghana Micro Report
28
Nonfarm income: participation and intensity
A Two-part model (TPM) is used for this purpose. The first part involves a probit model
predicting the probability of nonfarm income participation; the second part is a regression of
log nonfarm income on a set of covariates. A fixed-effects linear probability (FE-LP) model is
also estimated for the panel data first part regression to sweep out unobserved household
heterogeneity. For Afrint 3 where gender disaggregated income data was collected a bivariate
probit model is estimated in the first part, as the correlation coefficient between the error
terms is significantly different from zero at the 5% level.
The panel data results are in Table 21 while that based on the Afrint 3 sample can be found in
Table 29. The covariates include household demographic characteristics, household resource
endowments and spatial location. Both the penal data and Afrint 3 analysis results show the
probability of participation in nonfarm income to be decreasing with the number of ‘able’
household labour resources, and participation in nonfood cash crop production (particularly
commercial vegetables). We included a fallowing dummy in the models. A prior, the expected
effect of fallowing was ambiguous. If access to land is limiting then households could leave land
fallow while pursuing nonfarm work, bearing entry barriers. On the other hand, only land
abundant households could afford following. We observe that the probability of nonfarm
income participation is decreasing with land fallowing.
The probability of nonfarm participation is decreasing with farm size in the panel data
estimates but not throughout the entire distribution of the farm size distribution. Formal
education increases the probability of nonfarm work in the panel but in the Afrint 3 the
education effect shows up only in the gender-disaggregated estimates where the effect in
positive for men and negative for women.
The largest effect magnitudes on nonfarm participation probability in the panel estimates come
from time and spatial located effects: the probability of participation was approximately 20
percentage points higher in 2013; living in the Upper East villages increases the probability of
participation by between 15 to 23 percentage points compared with living in the ‘wealthiest’
village—Gyedi. In Afrint 3, credit access has a relatively high positive effect on the probability of
nonfarm participation. Also, the probability of nonfarm participation is significantly higher
among female than male farm managers by about 10 percentage points.
AFRINT 3: Ghana Micro Report
29
The advantages of the TPM over the more restrictive Tobit, for example, is evident as it can be
seen that the effect of the covariates on the probability of participation do not always carry
through to the level of participation. Examples from the panel data estimates are the number
of able household member as well as time and location effects. Particularly on village location,
it is observed that locating in the Upper East villages is associated with lower levels of nonfarm
income compared with locating in Gyedi, and indeed all other Eastern Region villages. The story
is not different for the Afrint 3 cross-sectional estimates. Also, from the Afrint 3 estimates
(Table 29), female farm managers receive approximately 25% lower nonfarm incomes
compared to their male counterparts, although participation rate is higher among females as
the first part regression indicates.
Finally, a bit more commentary on the gender-disaggregated estimates from the Afrint 3 data
(Table 29). Clearly, the determinants of the decision to participate in nonfarm work differ
across the genders and often in opposite directions; this is evidenced by the negative and
significant correlation coefficient, ρ, between the error terms in the two equations (ρ = –0.20).
This means that, overall, factors that tend to increase female participation decreases male
participation. For example, the probability of female participation is increasing with the
number of able workers, the proportion of dependents, and credit access but not so for male
participation. Similarly, the level of nonfarm income received by males is increasing with level
of education, proportion of dependants, and household ownership of sowing machine, but not
so for level of income received by females. The most important determinant of level of female
income is access to credit and spatial location.
Nonfarm income effects on farm outcomes
As mentioned earlier, because of the threat of measurement error in reported household
income data both the binary nonfarm income participation and the semi-continuous nonfarm
income variables are utilized as explanatory variables in separate equations. Table 44 contains
a summary of all the regression results. The full results of both the panel and Afrint 3 cross-
section estimates can be found at the end of this document.
AFRINT 3: Ghana Micro Report
30
We begin with the effect of nonfarm income participation on purchased input use. The a priori
expectation is a positive and significant coefficient on the nonfarm income variables. The
expected sign is always observed in the panel data results (Table 22) but there is no evidence
that either participation or level of nonfarm income significantly increases the probability of
purchased input use. For the Afrint 3 cross-section estimates (Tables 30 & 31) we observe a
negative and significant male nonfarm participation effect on the probability of purchased
input use when using the gender disaggregated variable and accounting for endogeneity via a
bivariate probit model under valid exclusion restrictions. The estimated magnitude of effect is
non-trivial: on average, nonfarm income participation by male household members lowers the
probability of purchase input use by approximately 19 percentage points.
The dependent variable in the second part of the model is log input expenditure. The panel
data results again show no significant effect of nonfarm income if endogeneity is allowed
through the household specific effect only (Table 22). Once nonfarm participation is allowed to
be correlated with the random error the expected positive and significant effect is detected.
The story is similar for the Afrint 3 estimates (Table 31) where we find additional evidence of a
significant positive association between purchased input expenditures and the level of female
nonfarm income. About 42% of households in the Afrint 3 sample have both male and female
nonfarm income earning members. We find that average purchased input expenditure of such
households is about 22% higher than the rest of the sample.
The next set of farm sector outcome variables of interest are improved seed adoption and
hired labour use. Our data contains information on improved seed adoption and a binary
indicator of whether or not the household regularly hires farm labour. Previous work (e.g.
Oseni and Winters, 2009; Stampini and Davis, 2009; Takahashi and Otsuka, 2009) shows that
households receiving nonfarm income spend significantly more on seeds and use more hired
labour. In our data (both the panel and Afrint 3), no evidence is found that nonfarm income
increases the probability of improved seed adoption (Tables 23 & 32). It is observed, however,
that having both male and female nonfarm income earners increase the average probability of
improved seed adoption by as much as 38 percentage points, after controlling for endogeneity.
We find mild evidence of a significant positive association between hiring farm labour and
nonfarm income: on average, nonfarm income participation is associated with a 10 percentage
points increases in the probability of hiring farm labour, ceteris paribus.
AFRINT 3: Ghana Micro Report
31
Next is nonfarm participation and income effect on staple crop output market participation.
The TPM is used. The first part predicts the probability of selling own-produced maize and rice,
and the second part the share of output put out for sale (Tables 24, 33 & 34). If nonfarm
income reduces the ‘safety first’ attitude that leads to self-provisioning of food requirements
then a significant positive effect is expected a priori. The results using the panel data show no
evidence that nonfarm income has an effect on either the probability of selling maize and rice
or the share of output sold by farm households. Some evidence is found in the cross-sectional
data but is not overwhelming, and only in the first part after disaggregating the nonfarm
income information by gender in the maize case. In the case of rice (Table 34) it is observed
that nonfarm participation lowers the probability of selling own-produced rice by 12
percentage points but conditional on selling, participation increases the proportion sold by
about 17%, a result that is consistent with findings in the literature (Reardon et al., 1994).
Disaggregating the income data shows that the positive effect on the share of rice sold comes
from male nonfarm participation. Although the first part result is contrary to a priori
expectation, this can be explained by the food insecurity situation in the Upper East where the
rice producing households are located. Households in this region would generally sell staples
only if they are severely constrained. Thus, where nonfarm income is available the probability
of selling is reduced, as observed from our results.
A similar argument as above has been made with respect to the possible effect of nonfarm
income on cash-crop production (Reardon et al., 1994; Huang et al., 2009). Tables 25 and 35
contain the panel and cross-sectional data estimates that seek to test this hypothesis. Because
the nonfood cash crop variable is captured as binary, the panel data estimates are from pooled
probit and fixed effects linear probability models while the Afrint 3 estimates are from probit
models. No evidence is found in the panel to support the hypothesis. The Afrint 3 estimates
provide some evidence of significant nonfarm income relationship, but only after using the
gender disaggregated information. It is observed that nonfarm income participation by males is
associated with a lower probability of participation in nonfood cash crop production while the
hypothesised positive relationship is related to female nonfarm participation. This result makes
sense if nonfood cash crop production relies on male family labour such that there is a trade-
off between spending their time working off and on farm.
AFRINT 3: Ghana Micro Report
32
Finally we explore the possible effect of nonfarm income on total cultivated area. One would
expect a significant positive effect of nonfarm participation on farm size if farming is viewed by
the household as the main occupation and as such see nonfarm income as an avenue for
relaxing the liquidity constraint that constrains farm expansion. If not, increased nonfarm
participation could serve as a catalyst for exiting the farm. From the panel data estimates
(Table 26) a positive and statistically significant coefficient is observed on the semi-continuous
nonfarm income variable after allowing endogeneity to operate through the random error. Yet,
the magnitude of effect is small: on average, a one dollar increase in nonfarm income is
expected to increase average cultivated area by less than 0.1%, all else held constant. In the
cross-sectional estimates from the Afrint 3 data, a negative sign is observed on the nonfarm
income participation variable, meaning that participation is associated with decreasing
cultivated area, on average (Table 36). Allowing participation in nonfarm income to be
endogenous, it is estimated that average total cultivated area among participants was
approximately 38% less than that of non-participants; under exogeneity, the difference is
approximately 15%. The gender disaggregated nonfarm income variables suggest that the
negative effect comes largely from female nonfarm participation. Also, households with both
male and female nonfarm income earners cultivate smaller farm sizes.
Distributional implications of nonfarm income
Attention is now turned to the effect of participation and level of nonfarm income on welfare
outcomes. Two household welfare indicators are utilised as described earlier: a welfare index
and food (in)security. Estimates of the panel data welfare index equation are in Table 27. They
are fixed and random effects estimates assuming the nonfarm income indicators are
exogenous (i.e. uncorrelated with the idiosyncratic error). Where this assumption is invalid, we
apply the two-step approaches suggested by Vella and Verbeek (see Vella and Verbeek, 1998;
Vella and Verbeek, 1999). The cross-section welfare index estimates are fit by OLS assuming
exogeneity of nonfarm income. Assuming endogeneity, a dummy endogenous regression
model and the two-step approach suggested by Vella (1993) is applied (Table 37). The food
insecurity status regressions (Tables 28 & 38) are pooled probit8 and fixed effects linear
probability model estimates for the panel data; probit is used for the cross-sectional data. In
8 Results from the random effects probit model show clearly that the panel level variance component was unimportant in the current setting, hence the pooled probit estimates.
AFRINT 3: Ghana Micro Report
33
accounting for possible endogeneity of nonfarm income and participation the bivariate probit
model is applied where the binary nonfarm income participation variable is the reduced form
equation. Where the semi-continuous nonfarm income level enters the model, the two
equations are estimated simultaneously using the multi-equation, multi- conditional mixed-
process estimators (Roodman, 2011).
As could be expected, the welfare index is increasing with both nonfarm income participation
and level in the panel data estimates. With the Afrint 3 estimates, we observe only a significant
positive association between the welfare index and the level of nonfarm income participation.
The difference in the average welfare index between nonfarm participants and non-
participants (in the panel estimates) is between three to four units—this value is about 26% of
the pooled median value of the welfare index. An increase of $100 in nonfarm income is
associated with an average increase of only 0.2 units in the average welfare index for the Afrint
3 estimates; the magnitude of effect in the panel is much higher (between 1.4 to 1.7 units).
To get a bette